Abstract A method for increasing the accuracy of an analog neural network which computers a sum-of-products between an input vector and a stored weight pattern is described. In one embodiment of the present invention, the method comprises initially training the network by programming the synapses with a certain weight pattern. The training may be carried out using any standard learning algorithm. Preferably, a back-propagation learning algorithm is employed. Next, the network is baked at an elevated temperature to effectuate a change in the weight pattern previously programmed during initial training. This change results from a charge redistribution which occurs within each of the synapses of the network. After baking, the network is then retrained to compensate for the change resulting from the charge redistribution. The baking and retraining steps may be successively repeated to increase the accuracy of the neural network to any desired level.
A method of increasing the accuracy of an analog circuit employing floating gate memory devices which are programmed by transferring charge onto a floating gate member insulated from a substrate by a tunnel oxide and from a control gate by a dielectric layer comprising an oxide-nitride-oxide stack, a dipole being formed in said dielectric layer upon programming of said floating gate memory devices, said method comprising the steps of: (a) programming said floating gate memory devices with a predetermined charge; (b) baking said circuit to accelerate the redistribution of charge within said dielectric, said redistribution causing a shift in the voltage threshold of said floating gate memory devices; and (c) reprogramming said floating gate memory devices to effectively cancel out said shift and restore said predetermined charge to said floating gate memory devices.
Claim 2
Independent Claim
133 words
A method of increasing the accuracy of an analog circuit employing floating gate memory devices which are programmed by transferring charge onto a floating gate member insulated from a substrate by a tunnel oxide and from a control gate by a dielectric layer comprising an oxide-nitride-oxide stack, a dipole being formed in said dielectric layer upon programming of said floating gate memory devices, said method comprising the steps of: (a) programming said floating gate memory devices with a predetermined charge; (b) baking said circuit to accelerate the redistribution of charge within said dielectric, said redistribution causing a shift in the voltage threshold of said floating gate memory devices; and (c) reprogramming said floating gate memory devices to effectively cancel out said shift and restore said predetermined charge to said floating gate memory devices.
Claim 3
Dependent Claim
24 words
The method of claim 1 further comprising the step of: (d) repeating steps (b) and (c) to further increase the accuracy of said circuit.
Claim 4
Dependent Claim
19 words
The method of claim 1 wherein said floating gate devices are both electrically programmable and electrical erasable memory devices.